LoRA Load Only (Yogurt Nodes)

"Load a LoRA without applying it" sounds like a contradiction, and it's the most useful idea in this pack. The normal LoraLoader loads the weights and slams them into your model and CLIP in one move - you never get to see the LoRA itself, let alone change it. YogurtLoadLoraOnly breaks that apart: it reads the LoRA file into memory as a manipulable object and stops there. No model patched, no CLIP touched, nothing sampled. Just the LoRA, sitting on the wire, waiting for you to do something to it.

This node is the entry point to the LoRA surgery suite in ComfyUI-YogurtNodes (yogurt7771's 150+ node all-in-one, MIT-licensed, auto-generated README, essentially unknown in the community). The pack's model section is where it gets interesting: merge, compress, scale - none of that is possible if you can't first load a LoRA without applying it.

How it works

One required input, lora_name - an enum of LoRA files from your ComfyUI/models/loras directory, so drop the file there and it appears. One output, lora, typed LORA. Important nuance: this LORA is the pack's own object type (a dict-like state of the LoRA's tensors and metadata), not the standard loader's output. You don't wire it into a normal LoraLoader - you wire it into the pack's operation nodes.

The flow looks like this:

YogurtLoadLoraOnly → [LoraMerge / LoraRankCompress / LoraScaleAlpha /
                       LoraScaleWeights / LoraSimpleAdd / LoraStatViewer]
                  → YogurtSaveLora  →  load the result with the normal loader

Or skip the file entirely and use the pack's Merge LoRA To Model to apply the processed LoRA directly to model + CLIP in-graph. Either way, this node is just the read step.

Why you'd bother

The whole point is baking changes in once instead of re-tuning every run. Want a LoRA at 0.7 strength without touching the slider every time? Scale its alpha, save it, and it comes out at 0.7 by default. Want two LoRAs merged into one file for a workflow that has to stay portable? Merge them and save. The operations are only as good as the object they start from, and that's what this node exists to produce.

Install

ComfyUI Manager, search ComfyUI-YogurtNodes, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

Restart and it's under "Yogurt Nodes". No model downloads - the LoRA files are yours, from models/loras. Dependencies are the pack's usual light set; the heavy AI-client libraries only load if you touch the LLM nodes.

One honest caution before you start operating on your files: keep backups. YogurtSaveLora writes a new safetensors file, so as long as you save to a new name you're safe - but the LoRA editing path here assumes standard .lora_down/.lora_up pairs, and if your LoRA is a DoRA or LoCon variant, check it with YogurtLoraStatViewer before running the rank-based operations on it.